Machine-learning algorithms to rethink the way we predict the weather
From predictive maintenance to autonomous decision-making, machine learning models are revolutionising many fields, but can they improve current weather forecasting models?
AtmoRep uses large-scale representation learning from artificial intelligence to determine a general description of the atmosphere’ dynamics.
Launched in March, the AtmoRep project aims to demonstrate this by applying a large-scale machine learning model to atmospheric dynamics, transferring the technology established for large linguistic models such as ChatGPT to earth system modelling.
The model has been trained on more than 5TB of Earth observation data collected over the past 40 years, and is showing promising preliminary results for near-real-time weather forecasting (nowcasting) and improving geographical accuracy (downscaling). This versatility sets AtmoRep apart from recent machine learning models for weather forecasting.
AtmoRep is the result of a collaboration between CERN, the Forschungszentrum Jülich and the Otto-von-Guericke University in Magdeburg.
CERN internal seed funding (CIPEA)
Machine learning for weather forecasting